{"id":"0748ffae-f882-4e5e-a83e-b28c9692617d","arxiv_id":"2508.19006","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"Pre-trained RNN models with global self-attention or sliding-window sparse attention deliver the highest value-weighted Sortino ratios (2.0 and 1.80) in a COVID-era backtest of 420 large-cap US stocks.","lead":"A single-author study tests pre-trained RNN models with several attention mechanisms for pricing 420 large US stocks, reporting that global self-attention and sliding window sparse attention give the best risk-adjusted backtest returns. The paper emphasizes causal masks to avoid future data leakage, a point it says prior finance attention models ignored.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Stock universe is selected using full test-period survival, making the reported Sortino ratios (2.0/1.80) non-tradable look-ahead results; a point-in-time universe check is required.","rationale":"The reader's weakest assumption is the same one I would flag; I sharpen it by noting it is not merely survivorship bias but look-ahead conditioning on the full test period. The paper's own motivation for the filter makes this worse, because it is an explicit design choice rather than an oversight. This concern is more load-bearing than the missing sparse-window size or lack of code, because it affects every portfolio result in Table 5. I would keep the reader's CONDITIONAL verdict: the causal-mask insight and model comparisons are informative, but the headline Sortino ratios should not be accepted until the point-in-time universe check is run. If the check preserves the ranking, the paper's central empirical claim survives on this dimension; if not, the paper needs major revision.","tokens_in":31819,"tokens_out":6379,"duration_ms":59752,"concrete_test":"Rerun the entire pipeline and the value-weighted backtest with a point-in-time universe: for each month from Jan 2013 through Dec 2022, include every stock that is listed and has a return at that date; when a stock stops trading, include its delisting return and remove it thereafter; do not impose any 'no missing values in testing period' filter. Compare the 2112 annualized Sortino ratio of self att and sparse att with Table 5. If the ratios fall below the corresponding RNN/LD benchmarks or below ~1.5, the headline downside-hedging claim is an artifact of the survivorship filter; if they remain ~2.0, the concern is empirically minor.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 3 defines the 420-stock universe as stocks with 'no missing value in the testing period and less than 50% of missing values during the training period,' with testing period 2013-2022. This conditions the backtest universe on information unavailable at the start of the OOS period: a stock that later experiences missing data or delisting is removed. Distressed/delisted firms are therefore excluded by construction. All value-weighted portfolio metrics in Table 5, including the headline annualized Sortino ratios of 2.00 (self att) and 1.80 (sparse att) in 2112, are computed on this filtered universe. The stated motivation ('too-big-to-fail', 'going concern') does not cure the look-ahead: a practitioner in 2013 cannot know which 420 stocks will have complete data through 2022. This is the load-bearing weakness because it directly concerns the central claim about downside hedging and absolute returns; no amount of model architecture refinement can fix a biased test set. Note that the causal-mask contribution is independent and remains valid.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes pre-trained RNN attention models — a global self-attention model and a sliding-window sparse attention model — for asset pricing on the top 420 large-cap US stocks. It compares them with additive attention, Luong's three attention variants, vanilla RNN, LSTM, and GRU, using 182 firm-characteristic-sorted portfolio factors. The models are evaluated by out-of-sample predictive accuracy (R2, MSE, alpha, Diebold-Mariano tests) and by equal- and value-weighted portfolio backtests over three periods: pre-COVID (1911), COVID (2112), and post-COVID (2212). The headline claim is that the global self-attention and sparse attention models achieve annualized Sortino ratios of 2.00 and 1.80, respectively, in the value-weighted COVID period (Table 5), and that the sparse attention model is more stable across market-cap sizes. The paper emphasizes the use of causal masks to avoid future-data leakage, which is a legitimate and under-appreciated point in the attention-based asset pricing literature.","tokens_in":32069,"tokens_out":3309,"duration_ms":32241,"significance":"If the empirical claims were robust, the paper would make a useful contribution by systematically comparing mainstream attention mechanisms in an asset pricing setting, by proposing simplified RNN-attention architectures that reduce overfitting risk, and by demonstrating that causal masking prevents a known look-ahead bias in attention models. The three-period design, including the COVID-19 crisis, is a strength, and the use of Diebold-Mariano tests for predictive accuracy is appropriate. However, the central backtesting claims — the Sortino ratios and the 'outperforms' statements — are not statistically supported and are based on a stock universe selected using full test-period survival. The missing specification of the sparse attention window size also undermines reproducibility. The causal-mask contribution is independent and appears sound, but the empirical asset pricing conclusions need substantial additional support before the headline claims can be accepted.","major_comments":[{"comment":"The stock universe is selected using information unavailable at the start of the out-of-sample period. The 420 stocks are required to have 'no missing value in the testing period and less than 50% of missing values during the training period,' with the testing period extending to December 2022. This removes any stock that was delisted or had data interruptions during 2013–2022, thereby excluding distressed and failed firms by construction. The value-weighted portfolio results in Table 5, including the headline annualized Sortino ratios of 2.00 (self att) and 1.80 (sparse att) in the 2112 period, are computed on this survivorship-biased universe. The stated 'too-big-to-fail' and 'going concern' motivation does not cure the look-ahead: a practitioner in 2013 cannot know which 420 stocks will have complete data through 2022. A point-in-time universe construction, where eligibility is based","section":"Section 3, Tables 4–5"},{"comment":"No confidence intervals or statistical significance tests are reported for the differences in Sharpe ratios, Sortino ratios, or annualized returns across models. The Diebold-Mariano tests in Section 5.1 (Table 3) assess differences in absolute prediction errors, not differences in portfolio performance metrics. Consequently, statements such as 'the self att model outperforms all alternative models' and 'the sparse att model performs more stably' rest on point estimates that could easily be within sampling variation, especially given only 83–120 monthly observations per period. The paper should provide bootstrap or HAC-based confidence intervals for the Sharpe and Sortino ratio differences, or otherwise quantify the uncertainty around the rankings. Without this, the central empirical claim is not statistically secured.","section":"Section 5.2, Tables 4–5"},{"comment":"The sliding-window sparse attention model is a proposed core contribution, but the attention window size w is never specified. Equation (39) defines the attention set S_t = {s | max(0, t−w) ≤ s ≤ t}, but no value is given in the model description, data section, or empirical results. This hyperparameter directly controls the trade-off between computational cost and long-term dependency capture and affects the reported performance. The manuscript also omits the L1 regularization coefficient λ, the early-stopping patience parameter, and the exact latent dimension of the MLP autoencoder (70% of 182 is 127.4, which is not an integer). These details are necessary for reproducibility and for assessing the sensitivity of the results to model choices.","section":"Section 4.3.4 and Section 5"}],"minor_comments":[{"comment":"Numerous typos and inconsistent terminology distract the reader: 'annulized' for 'annualized', 'BHV' vs 'BHE' for buy-and-hold, 'roling' for 'rolling', 'prevelent' for 'prevalent', 'cancatenate' for 'concatenate', 'pertaining' for 'pre-training', and 'Univiersity' in the author affiliation. The paper would benefit from careful proofreading.","section":"Throughout"},{"comment":"The text refers to 'Equation (65)' for the linear output of the RNN attention models, but the equation is actually numbered (41). The equation numbering throughout Section 4 appears shifted; please verify all cross-references.","section":"Section 4.4"},{"comment":"The MLP autoencoder latent dimension is described as '70% of the original input.' With 182 input factors, this implies a non-integer number of latent units. Please state the exact integer dimension used and how it was determined.","section":"Section 4.1"},{"comment":"The period labels '1911', '2112', and '2212' are not immediately intuitive; although the text explains them, a sentence in Table 1 or the caption would improve clarity. Also, 'Observation No.' for the in-sample period is 672, but it is unclear whether the validation window is included in this count.","section":"Section 3, Table 1"},{"comment":"In the maximum drawdown formula, the notation Cmax(t) and C(t) is introduced but the definition of 'highest value during time t' is imprecise. Please clarify whether this is the running maximum of the cumulative return path.","section":"Section 5, Eq. (79)"}],"recommendation":"major_revision","confidential_remarks":"The paper reads like a thesis chapter in places, with a large number of self-citations and a somewhat unstructured literature review. The core technical contribution — causal masking in attention mechanisms for asset pricing — is sound and worth preserving. However, the empirical claims currently rely on a survivorship-biased universe and lack statistical inference for portfolio metrics. These are fixable with additional experiments and robustness checks, so major revision seems appropriate rather than rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is a solid empirical comparison of attention-augmented RNNs for asset pricing, with a genuinely useful causal-mask point, but the headline Sortino ratios rest on a survivorship-biased universe and lack statistical support. The central claim is not established as stated.\n\nWhat's new and good: the paper systematically compares nine models (vanilla RNN, LSTM, GRU, additive, three Luong variants, global self-attention, sliding-window sparse attention) on 420 large-cap US stocks across three periods around COVID. It uses public data (Chen/Zimmermann factors, WRDS, French data). The enforced causal masks in attention are a legitimate correction: several recent transformer-based asset pricing papers do ignore the single-direction nature of time series, and the author is right to flag that. The finding that the vanilla RNN has the highest OOS R-squared while the attention models do better on portfolio metrics is reported honestly, not buried. The paper also acknowledges limitations (pre-selected factors, positive alpha).\n\nWhere it's soft: the big one is the universe selection. Section 3 requires no missing values in the testing period and less than 50% missing in training. That conditions on survival through 2022, so the backtest is not tradable. A practitioner in 2013 cannot know which stocks will have complete data a decade later. The stated 'going concern' justification doesn't cure the look-ahead. This is load-bearing: the 2.0 and 1.80 Sortino ratios in Table 5 are computed on this filtered universe, so the downside-hedging claim is not reliable.\n\nOther issues are secondary but real. There are no confidence intervals or significance tests for the Sharpe/Sortino differences, so the model rankings could be noise. The sparse attention window size is never specified, which matters for a method whose whole point is the window. No code or data is provided, making reproduction harder. The 'first paper' novelty claim is overstated given prior attention-based asset pricing work, though the specific combination may be new.\n\nThe causal-mask point is independent of the backtest problem and remains valid. The paper is not incoherent, and the internal logic is fine; it's the test set that's biased.\n\nWho it's for: readers interested in time-series attention for finance will get something from the causal-mask discussion and the systematic comparison. But the headline result should not be taken at face value. I'd send it to peer review because it has a legitimate methodological point and a substantial empirical study, but it needs major revision: a point-in-time universe, statistical tests on portfolio metrics, and the window size reported. If those are addressed, the comparison could be useful.","headline":"Useful causal-mask correction and systematic comparison, but the headline Sortino ratios are built on a survivorship-biased universe and lack statistical support; central claim needs major revision.","tokens_in":32556,"tokens_out":2680,"would_cite":false,"duration_ms":24364,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper claims that pretrained RNN models with global self-attention or sliding-window sparse attention, fitted with causal masks, deliver strong absolute returns and downside-risk hedging in long-only value-weighted backtests on 420 larg","keywords":["attention mechanisms","asset pricing","recurrent neural networks","factor models","self-attention","sparse attention","causal mask","Sortino ratio"],"falsifier":"Run the same backtest on a point-in-time universe that includes stocks delisted or with missing data during 2013–2022, using delisting returns and no survivorship filter; if the global self-attention and sparse-attention models no longer beat buy-and-hold on Sortino ratio during the COVID period, the central performance claim fails. A second check: remove the causal masks and retrain on identical inputs; if Sortino ratios do not deteriorate, the paper's distinct claim about future-information leakage is not what drives the results.","tokens_in":31671,"feed_emoji":"📈","tokens_out":5592,"duration_ms":49344,"temperature":0.7,"pith_summary":"This paper tries to establish whether attention mechanisms borrowed from natural-language processing improve machine-learning asset pricing, and specifically whether two attention variants mounted on a pretrained RNN—global self-attention and sliding-window sparse attention—beat vanilla RNN, LSTM, GRU, and other attention benchmarks. It reports that in long-only value-weighted backtests with 50 basis-point transaction costs, the global self-attention model and the sliding-window sparse attention model achieve annualized Sortino ratios of 2.0 and 1.80 in the period containing COVID-19, outperforming buy-and-hold and most benchmarks on downside risk. The models also produce positive out-of-sample residual alphas across all periods, which the paper interprets as evidence of persistent mispricing rather than model misspecification. A distinctive technical claim is that causal masks prevent future-information leakage that prior attention-based asset-pricing models ignored.","feed_headline":"Causal-masked attention models post Sortino 2.0 in COVID test","feed_subtitle":"RNN plus global or sparse attention beats buy-and-hold on downside risk for 420 large-cap US stocks.","key_machinery":"The central object is the pre-trained RNN attention model: an MLP autoencoder compresses 182 factor-portfolio inputs to 70% of their dimension and imputes missing values; a two-layer tanh RNN with 64 then 32 neurons encodes the time series; an attention layer—either global self-attention (scaled dot-product of query, key, and value projections) or sliding-window sparse attention (the same projections but only over the most recent w time steps)—reweights past RNN hidden states; and a causal mask forces j ≤ t in the attention scores. A linear layer maps the attention output to next-month excess return. The causal mask is the load-bearing modification: it makes the attention weights respect tim","core_discovery":"On the paper's own terms, the discovery is that attention mechanisms designed for long text—global self-attention from the Transformer and sliding-window sparse attention from Longformer—can be mounted on a two-layer RNN, pretrained with an MLP autoencoder on firm-characteristic-sorted portfolio factors, and produce tradable long-only signals for 420 large-cap US stocks. The argument is carried by a causal mask that restricts each time step's attention to past time steps, addressing look-ahead bias the paper attributes to earlier attention-based asset-pricing work. In the value-weighted backtest over January 2013 to December 2021, the period that includes COVID-19, the global self-attention","pith_inferences":["Editorial inference: the survivorship-biased sample—only stocks with no missing testing-period values—likely inflates the reported Sortino ratios; a point-in-time universe including delisted stocks would probably lower them, though the relative ranking of the two attention models might survive.","Editorial inference: the trend-following signal is defined by agreement between predicted and realized signs; used as described, that agreement criterion is only known ex post, so the backtest may embed a mild look-ahead even with causal attention masks; a tradable version should act only on predicted signs.","Editorial inference: the same architecture could be transferred to higher-frequency or international data, where sparse attention's linear cost in window size would be a practical advantage; a natural test is whether the 2.0 Sortino persists out of sample after including delistings.","Editorial inference: the paper's comparison to unmasked Transformer models suggests a direct ablation—masked versus unmasked attention on identical RNN inputs—would quantify how much of the reported performance is due to the causal mask alone."],"forward_implications":["If the claims hold, long-only strategies built on causal-masked RNN attention signals are viable for large-cap US equities, delivering higher Sortino ratios and lower drawdowns than buy-and-hold in volatile markets.","Attention-based asset-pricing studies that omit causal masks should be treated with suspicion, because their reported performance may include future information.","Sliding-window sparse attention is the safer default for practitioners because its performance is less sensitive to market-cap weighting than global self-attention.","The persistent positive residual alphas imply that the 182-factor set does not fully price the cross-section, leaving room for factor discovery.","Value weighting tends to compress model differences and improve risk-adjusted returns, meaning model rankings depend on portfolio construction."],"supporting_citations":[{"why":"Supplies the global self-attention mechanism that the paper adapts to an RNN frame with a causal mask.","marker":"[11]"},{"why":"Supplies the sliding-window sparse attention mechanism that the paper abstracts as its other proposed model.","marker":"[12]"},{"why":"Provides the large-scale ML asset-pricing benchmark and evaluation setup the paper builds on and compares against.","marker":"[3]"},{"why":"The closest prior work on RNN attention models for asset pricing, which the paper extends by adding causal masks and new attention variants.","marker":"[6]"},{"why":"Introduces autoencoder-based asset pricing, the source of the MLP autoencoder pretraining idea.","marker":"[20]"},{"why":"Supplies the 182 firm characteristic-sorted portfolio factors used as original inputs.","marker":"[5]"},{"why":"Defines the three Luong attention mechanisms the paper implements as benchmark attention models.","marker":"[31]"},{"why":"A recent attention-based asset-pricing model the paper critiques for not enforcing causal masks.","marker":"[13]"}],"fun_headline_variants":["Attention models hit Sortino 2.0 in COVID backtest","Causal-masked attention beats buy-and-hold on risk","Global and sparse attention hedge downside for large caps","RNN-attention models tackle look-ahead bias in pricing"],"cache_read_input_tokens":2688,"weakest_assumption_plain":"The results are computed on 420 large-cap stocks selected because they had no missing values in the test period and few missing values earlier, which excludes stocks that were delisted or stopped trading; if those excluded stocks are part of the real opportunity set, the reported downside-risk performance may not be achievable.","fun_headline_variants_meta":{"raw":{"variants":["Attention models hit Sortino 2.0 in COVID backtest","Causal-masked attention beats buy-and-hold on risk","Global and sparse attention hedge downside for large caps","RNN-attention models tackle look-ahead bias in pricing"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000446,"raw_usage":{"total_tokens":2139,"prompt_tokens":840,"completion_tokens":1299,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":584,"completion_tokens_details":{"reasoning_tokens":1230}},"tokens_in":584,"tokens_out":1299,"duration_ms":8887,"temperature":1.0,"reasoning_tokens":1230,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T16:01:28.768601+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the same backtest on a point-in-time universe that includes stocks delisted or with missing data during 2013–2022, using delisting returns and no survivorship filter; if the global self-attention and sparse-attention models no longer beat buy-and-hold on Sortino ratio during the COVID period, the central performance claim fails. A second check: remove the causal masks and retrain on identical inputs; if Sortino ratios do not deteriorate, the paper's distinct claim about future-information leakage is not what drives the results.","supporting_citations":[{"cited_title":"Review of Financial Studies 33(5), 2223–2273 (2020) https://doi.org/10.1093/ rfs/hhaa009","cited_arxiv_id":null,"evidence_quote":"Provides the large-scale ML asset-pricing benchmark and evaluation setup the paper builds on and compares against."},{"cited_title":"Available at SSRN 4999011 (2024)","cited_arxiv_id":null,"evidence_quote":"The closest prior work on RNN attention models for asset pricing, which the paper extends by adding causal masks and new attention variants."},{"cited_title":"Journal of Econometrics 222(1), 429–450 (2021)","cited_arxiv_id":null,"evidence_quote":"Introduces autoencoder-based asset pricing, the source of the MLP autoencoder pretraining idea."},{"cited_title":"Centre for Financial Research (CFR), working paper (2020)","cited_arxiv_id":null,"evidence_quote":"Supplies the 182 firm characteristic-sorted portfolio factors used as original inputs."},{"cited_title":"Technical report, National Bureau of Economic Research (2025)","cited_arxiv_id":null,"evidence_quote":"A recent attention-based asset-pricing model the paper critiques for not enforcing causal masks."}],"review_version":1}